Ice-water Classification in the Barents Sea from Sentinel-1 EW SLC Images

Abstract

The focus of this thesis is to do ice-water classification on Sentinel-1 EW SLC imagery using an unsupervised Mixture of Gaussian segmentation algorithm. The classes are automatically labelled as sea ice or water based on the slope of the observed brightness decay from near- to far-range, which is different for sea ice and water. The aim of the thesis is twofold. In the first part, the ability of seven features to separate between ice and water are evaluated. It turned out that a combination of geometric brightness, cross-polarization ratio and relative kurtosis gave highest accuracy - 99.29 %. In the second part, the goal was to find the highest achievable, reliable resolution on the classified ice-water maps. The purpose is to find out how little it is possible to multi-look and still achieve high enough radiometric resolution to discriminate ice and water in a satisfactory manner. It turned out that a resolution of 46x43 meter with moderate accuracy is obtainable if swath 1 is omitted. This require SLC imagery. Alternatively, 93x87m resolution with better ice-water separation and higher accuracy can be obtained. The classified image can be used to derive other ice information products, such as ice concentration. This was demonstrated both with low and high resolution, and compared with hand drawn ice concentration maps from the Meteorological Institute.